Decorative 3D object: an iridescent sphere whose surface slowly flows, surrounded by a halo of 163 points — one for each public repository and starred reference, with the 46 verified as his own work drawn in the accent color. As the page scrolls the sphere moves beside each section and changes form; in the open-source section the halo opens out. Every value it encodes is written in the page text.
Mansurbek Satarov
Ph.D. researcher in trustworthy AI, and the engineer who ships it.
I study how behavior forms inside a model while it trains, and build the systems that keep it trustworthy once it ships.
Every claim, checked
Nothing on this site is asserted. A claim is marked verified only where a public, machine-checkable artifact backs it; everything else is labeled self-reported, in plain sight.
Claims on this site
0 of 5 checked
- WorkVerified20 of 22 case studies resolve to a public repository
- Open sourceVerified46 repositories classified as own work from commit data
- PlaygroundVerifiedFive playable systems, each with a pure core under unit test
- ResearchVerifiedResearch direction stated; individual work not yet citable
- Self-reportedSelf-reported2 projects are résumé-declared with no public repository
Self-reported is not a failure state. It marks work declared on a résumé that no public repository can confirm — kept on this list so you know exactly which claims you can check yourself.
19 research areas, four questions
Research →How model behavior forms during training, how to keep it predictable once it has, what the model is made of, and what changes when it starts using tools.
Training & learning dynamics
Safety, security & governance
Models, modality & efficiency
Agentic & retrieval systems
RLHF & preference optimization
How preference data and policy optimization shape model behavior.
Training & learning dynamicsHow behavior forms while a model is being trained, and how to trace it back.
Selected work
All 22 projects →Research systems, agentic pipelines and production ML, each with the architecture that made it work and the results it actually produced.
Multi-Agent Deep Research
A LangGraph supervisor/worker research workflow that plans, gathers evidence, extracts structured findings, and writes sourced reports.Python · LangGraph · OpenAI API
CompleteRetail Intelligence Platform
End-to-end Azure retail analytics: customer lifetime value, churn risk, basket associations, household search, and cached dashboards.FastAPI · React · Vite
CompleteRAG Assistant
A modular retrieval-augmented generation pipeline: ingestion, chunking, embeddings, ChromaDB vector search, and a conversational interface.Python · Flask · ChromaDB
CompleteDinoMind Evolution
NEAT neuroevolution learns to play a dino runner — neural networks evolve generation by generation to dodge obstacles.Python · Pygame · NEAT-Python
CompleteDocuParse
High-performance PDF/EPUB/MOBI → structured Markdown conversion with AI layout detection, LaTeX equation handling, and multilingual support.Python · OCR · Layout-detection models
CompleteDigitRecognizer
A neural network handwritten from scratch in NumPy — forward propagation, backprop, and gradient descent on MNIST, no frameworks.Python · NumPy · Matplotlib
Complete
163 points — one for every public repository and starred reference. The 46 in blue are his own work, verified from commit history.
Open source, honestly labeled
Open source →Every public repository, labeled by a relationship computed from commit data. Forks and stars are listed as exploration and reading — never counted as contribution.
- Own work
- 46
- Forks kept for study
- 39
- Saved references
- 78
- Technologies used
- 59
Own work by domain
- Web20
- Data science7
- Computer vision6
- LLM & agents4
- Cloud3
- Systems2
- Games2
- Other1
- Machine learning1
Bars count repositories the classifier resolved to his own commit history. Forks and stars are excluded from every bar here.
Label definitions
- Built by me
- Owned, not a fork, with a substantial own-commit history.
- Co-built
- Owned, with two or more active co-contributors.
- Contributed
- Someone else's repository, with merged pull requests.
- Forked & explored
- A fork kept for study. Counts as reading, not work.
- Studied / used
- Depended on or read closely, without commits.
- Saved resource
- Starred for reference. Zero contribution claimed.
Things you can play
Playground →Five systems that run in your browser. Each has a pure, deterministic core under unit test — the code the tests replay is the code you play.
DinoMind Arena
Race a visualized AI runner over the identical seeded course.
Re-Entry Corridor
Hold the descent inside the corridor on a fixed propellant budget.
Go / No-Go Drill
Read four gauges and call the burn before the clock runs out.
Neural Tic-Tac-Toe
Exact alpha-beta on 3×3, depth-limited search on 3×3×3.
Typing Challenge
Machine-learning terms and real code, scored on live WPM and accuracy.
Let’s build something trustworthy
Research collaboration, Ph.D.-adjacent work, and engineering that has to hold up. Replies usually within two days.
Open to research collaborations and internships.